Dropout from psychotherapy is common, often occurring early in treatment, but predicting which patients will leave care remains difficult. In this retrospective cohort study of 31,146 patients receiving individual outpatient therapy through an employer-sponsored digital mental health platform, we examined whether language from first-session transcripts could be used to predict subsequent dropout (failure to attend a fourth session within 1 year of the first session). We combined clinical, demographic, and utilization variables with features derived from transcripts using natural language processing and large language models. Models incorporating transcript-derived features improved prediction compared with pre-session data alone (area under the curve, 0.642 vs 0.569). Provider behaviors, particularly whether follow-up care was clearly discussed, were strongly associated with retention, while patient-level constructs such as motivation, ambivalence, and therapeutic alliance were not consistently predictive. Although overall discrimination was modest, transcript-derived features improved identification of high-risk patients. These findings suggest that early-session dynamics contain actionable signals of engagement and may support scalable approaches to improving retention in mental health care.
School characteristics are associated with student mental health cross-sectionally, but little is known about effects of school factors over time. This study assessed whether school-level (regional, administrative, and student characteristics), staff-level (school culture and promotion of wellbeing), and student-level factors (school climate and connectedness) were associated with students’ distress and wellbeing cross-sectionally and 2 years later. The sample consisted of N = 1548 adolescents in Year 8 (Mage = 13.9) from 35 schools in Australia. Results showed that greater student-level socioeconomic advantage was associated with better wellbeing cross-sectionally. Higher levels of student-rated school connectedness and climate were associated with greater wellbeing and lower distress cross-sectionally. Longitudinally, only higher student-rated school connectedness was associated with higher wellbeing two years later. No school-level factors were associated with student wellbeing or distress, cross-sectionally or two years later. Findings highlight the importance of students’ sense of connectedness with their school community on wellbeing over time.
This observational survey study investigated the relationship between prior ayahuasca experience and mental health and mindset outcomes. Participants planning to attend an ayahuasca ceremony completed a survey before their upcoming ayahuasca use. Measures included substance use, depression, anxiety, stress, spiritual well-being, aberrant salience, and metaphysical beliefs. Participants were categorized into three groups according to their prior ayahuasca experience: naïve, low/moderate, and high-experience users. High-experience users reported significantly lower alcohol use, depression, stress, and anxiety, as well as higher spiritual well-being and stronger metaphysical beliefs. Aberrant salience scores were higher in both low- and high-experience groups compared to naïve users, with no differences observed between the two experienced groups. Tobacco use differences were modest, and other cognitive measures showed smaller effects. These findings suggest that greater prior experience with ayahuasca is associated with improved psychological and spiritual well-being and stronger metaphysical beliefs, with more limited effects on substance use and cognitive variables, highlighting areas for further investigation into its potential therapeutic applications.
Early detection of mental health risk in non-clinical, non-help-seeking populations is a persistent challenge. We investigate large language models (LLMs) for identifying individuals at high risk of mental illness using standardized speech tasks. LLMs perform poorly on spontaneous, non-declarative speech due to limited self-disclosure and the lack of suitable training data. To address this, we propose the Latent Emotion Proxy for Mental Health Detection with AI-integrated Decision-making (LEMon-AID), a multi-step prompting framework that leverages LLMs’ emotion profiling capabilities, integrated with output from machine learning models. This enables LLMs to link emotional features with mental health labels, significantly improving AUC from 0.55 to 0.67. While more improvement is needed, this novel hybrid design introduces layered reasoning that enhances performance while improving explainability and protecting data privacy. Our results demonstrate that emotion-aware prompting, augmented with cross-model integration, improves early risk detection and LLM interpretability.
